China Telecom AI · 2026-09-16 · major
Xing4.0-29B-A4B — China Telecom's agent model trained on Ascend NPUs
Xing4.0-29B-A4B is an Apache-2.0 mixture-of-experts model with 29B parameters and 4B active per token. It scores 75.00 on SWE-bench Verified and 57.50 on Terminal-Bench 2.1, and was trained entirely on Ascend 910C chips.
A 29B open-weight agent model from China Telecom, trained end to end on Chinese Ascend silicon.
Quick facts
| Maker | China Telecom AI |
|---|---|
| Parameters | 29B total, 4B active |
| Context window | 256K, extensible to 512K |
| License | Apache-2.0 |
| Architecture | mHC + MLA + MTP, 64 routed experts |
| Training hardware | Ascend 910C with MindSpore |
| Formats | Transformers, FP8, GGUF |
Benchmarks
What is it?
Xing4.0-29B-A4B is the first model in the Xing series, which China Telecom Artificial Intelligence Technology renamed from TeleChat. It holds 29B total parameters but activates only 4B per token, and reads 256K tokens of context natively, extensible to 512K. The weights are Apache-2.0 and ship in Transformers, FP8 and GGUF formats.
How does it work?
The architecture combines mHC, MLA attention and MTP, with 64 routed experts and one shared expert across 40 layers. Training ran on Ascend 910C clusters through MindSpore and MindFormers rather than NVIDIA hardware, and the team reports roughly 96% higher training throughput than the out-of-the-box setup after fine-grained MoE communication tuning, selective recomputation, DVM graph-operator fusion and custom Ascend C fused operators.
Why does it matter?
Agent work is where Xing4.0-29B-A4B is aimed, and the scores back that: 75.00 on SWE-bench Verified beats Gemma4-26B-A4B by 22 points at a similar active size, and 57.50 on Terminal-Bench 2.1 leads both listed rivals. The model was format-aligned for OpenCode, Claude Code, OpenClaw and Hermes, so teams that want a self-hosted coding agent can drop it into a harness they already run. It also shows a full frontier-scale training run finishing on domestic Chinese accelerators.
Who is it for?
teams self-hosting coding agents
Frequently asked questions
- How does Xing4.0-29B-A4B compare to Qwen3.6-35B-A3B?
- Xing4.0-29B-A4B trades places with Qwen3.6-35B-A3B depending on the task. Qwen edges it on SWE-bench Verified (76.00 to 75.00), AIME2026 (92.70 to 90.00) and Tau3-Bench (67.20 to 64.63), while Xing4.0-29B-A4B leads on Terminal-Bench 2.1 (57.50 to 51.50), Claw-Eval (76.55 to 74.54) and DeepresearchBII (60.80 to 59.70).
- What hardware was Xing4.0-29B-A4B trained on?
- Xing4.0-29B-A4B ran entirely on Huawei Ascend 910C clusters using the MindSpore and MindFormers stack, with no NVIDIA hardware in the loop. China Telecom says it is the first model of this scale trained that way, and that co-optimisation work — MoE communication tuning, selective recomputation, DVM graph-operator fusion and Ascend C fused operators for mHC — lifted training throughput by about 96%.
- Which agent frameworks and servers support Xing4.0-29B-A4B?
- Xing4.0-29B-A4B was format-aligned for the OpenCode, Claude Code, OpenClaw and Hermes agent frameworks. For serving, China Telecom lists vLLM, SGLang and KTransformers, and the model is exposed through an OpenAI-compatible chat completions API with a thinking toggle. Fine-tuning is supported through LLaMA-Factory and MindFormers.
- How were the SWE-bench and Terminal-Bench numbers measured?
- China Telecom reports SWE-bench Verified for Xing4.0-29B-A4B using the SWE-agent harness with a 210K context window at temperature 1.0. Terminal-Bench 2.1 used the terminus-2 harness at temperature 0.8 with a 64K output cap, a 24-hour timeout, and the score averaged over three runs. The model card lists the exact sampling settings for every benchmark.
Try it
Model id XingChen-AGI/Xing4.0-29B-A4B — serves on vLLM, SGLang or KTransformers